AnalysisNov 12, 20266 min read8 sources

Rubin beats GB300 by about 2x per dollar on agentic serving, not the 30x to 67x headline figures

SemiAnalysis's AgentX comparison shows Vera Rubin NVL72 with a 1.8x to 2.0x lead in tokens per dollar at 100 to 120 tok/s/user. The double-digit multiples appear only at high interactivity, where GB300 falls off. All Rubin numbers are previews on early software.

In one minute

SemiAnalysis's AgentX comparison shows Vera Rubin NVL72 with a 1.8x to 2.0x lead in tokens per dollar at 100 to 120 tok/s/user. The double-digit multiples appear only at high interactivity, where GB300 falls off.

InferenceX is SemiAnalysis's open-source inference benchmark, formerly called InferenceMAX. AgentX is its agentic-coding scenario. It replays recorded Claude Code sessions, with long shared prefixes, subagents and tool pauses, and the median request has about 142k input tokens. Since August 2026, NVIDIA and SemiAnalysis have each published Vera Rubin NVL72 versus GB300 NVL72 numbers on AgentX, with headlines ranging from 30x to 67x. NVIDIA's Vera Rubin NVL72 product page contains none of these AgentX numbers. It compares Rubin with GB200 on a different workload.

The question for builders is what these ratios actually measure. Rubin costs more per GPU-hour than GB300, so the useful number is tokens per dollar at the interactivity you actually serve, not a peak multiple.

What we checked

  • Can the original AgentX results and methodology be checked against a public source, including the cost assumptions and workload configurations behind the Vera Rubin-to-GB300 ratios? confirmed. SemiAnalysis's InferenceX page, dated October 19, 2026, publishes the per-model tables, the $/GPU-hr assumptions, the power figures, the engines and precisions, and the interpolation method. The AgentX methodology page and the open-source repository document the workload. Nobody has independently reproduced the results, and the Rubin numbers are labeled preview. [1][5][7]
  • Do the headline 30x (NVIDIA) and 67x (SemiAnalysis) figures represent typical agentic serving? refuted. Both figures come from single high-interactivity points: 160 and 170 tok/s/user, against specific GB300 configurations. At 60 to 120 tok/s/user, both sources show roughly 1.4x to 3x. [2][3][1]

What we found

  • SemiAnalysis's October 19 InferenceX comparison covers four models: DeepSeek-V4-Pro, MiniMax-M3, Qwen3.8-Flash-Next and GLM-5.2. At 100 tok/s/user (120 for GLM-5.2), it finds a 1.8x to 2.0x Rubin lead in tokens per $1 of TCO. The cost inputs are hyperscaler owning costs, and the page says a different cost or power figure would rescale every ratio. At peak throughput the lead shrinks to about 1.05x to 1.17x. The largest ratios, about 8x to 10x, sit at high interactivity. [1]
  • NVIDIA's August 24 technical blog claims up to 30x higher throughput per megawatt against GB300 on AgentX DeepSeek V4-Pro at 160 tok/s/user. NVIDIA ran that test itself on the SemiAnalysis workload, and the blog says the results were pending SemiAnalysis review. The blog's own figure shows the advantage at 2x at 110 tok/s/user and 10x at 130. [2]
  • SemiAnalysis's September 14 newsletter has the 67x headline. That figure is one 170 TPS point against GB300 Dynamo TRT-LLM under owning costs, and the article itself puts the gain at 1.4x to 3x across 60 to 100 TPS, where it says most providers would serve. The article also notes that the Rubin software was pre-release. [3][4]
  • The AgentX methodology page describes the v1.0 corpus: 393 sessions, built June 21, 2026. Each run starts at a seeded point in the session, completes a warmup, and then measures a one-hour profiling window. The page says AgentX does not evaluate answer quality, and it models speculative decoding with acceptance lengths measured elsewhere. [5]
  • NVIDIA's Vera Rubin NVL72 product page does not mention AgentX or GB300. Its inference claims are against GB200 NVL72, on Kimi-K2-Thinking at 32K/8K input/output sequence lengths. [6]
  • Who published the dashboard is not entirely clear. The InferenceX repository says only it holds official results. inferencex.com, the dashboard the README links to, redirects to inferencex.semianalysis.com, yet that site's pages show text describing them as an unaffiliated copy. This is probably a template string from the open-source frontend, but that has not been confirmed. [7][8]

What it means

For capacity planning, assume Rubin gives about 2x more tokens per dollar than GB300 on long-context agentic coding at 100 to 120 tok/s/user. The 30x and 67x multiples only apply if your product needs 160 tok/s/user or more, where GB300's frontier collapses. Every per-dollar ratio scales with the assumed $/GPU-hr. Rental pricing differs a lot from owning cost: SemiAnalysis cites about $8.5 versus $5 per chip-hour on three-year commits. So redo the math with your actual quotes.

The results are previews. Rubin software is early, the engines differ between systems, some GB300 runs are older, and the power figures are planning numbers rather than measurements. Because AgentX is open source (Apache 2.0, and runnable on upstream AIPerf), the most reliable step is to replay it on capacity you can actually rent before you commit.

Competing readings

  • At the interactivity where most agentic traffic is served today, Rubin's real AgentX advantage over GB300 is about 2x per dollar. The 30x and 67x headlines come from points where GB300's throughput collapses at high interactivity (likely, about 65%). What would tell it apart: SemiAnalysis's October comparison gives 1.8x to 2.0x per dollar at 100 to 120 tok/s/user on four models [1]; The 67x figure comes from 170 TPS against a single GB300 TRT-LLM configuration. NVIDIA's 30x figure comes from 160 tok/s/user [2][3].
  • The large multiples are the right signal, because agentic coding will move to high-interactivity serving, where GB300 falls off and Rubin keeps its throughput (unlikely, about 25%). What would tell it apart: NVIDIA presents the 30x figure at 160 tok/s/user as the agentic headline [2]; Rubin software is pre-release, so these gaps could widen [3][1].
  • The ratios mostly reflect differences in software stack and cost assumptions, and Rubin's underlying advantage at matched software is well under 1.5x (very unlikely, about 10%). What would tell it apart: The comparisons rely on assumed $/GPU-hr figures and planning power figures, and the engines differ between the systems [1]; NVIDIA measured the 30x result itself [2].

Base rate

Of 21 earlier research stories about organisations, products or projects in the New Runtime archive over the last 12 months that are at least 3 months old, 9 (43%) were followed within 3 months by a release story on the same company or product line. Of 21 comparable research stories in New Runtime's archive, 9 were followed within three months by a release story on the same product line. With Rubin already shipping and more InferenceX previews promised, a follow-up is plausible but not the default outcome.

What would change this view

  • A non-preview InferenceX AgentX run with mature vLLM, SGLang and TRT-LLM stacks on both systems shows Rubin far above 2x per dollar at 100 tok/s/user (model-listing signal): The roughly 2x mainstream gap was an artifact of immature software, and the larger multiples hold at normal operating points.
  • Rubin rental or owning cost moves well away from SemiAnalysis's $3.61/GPU/hr, or GB300 cost moves away from $2.31/GPU/hr (pricing signal): Every per-dollar ratio scales with the change, so the 1.8x to 2.0x figure would shift.
  • Independent operators publish AgentX submissions on Rubin that diverge sharply from the NVIDIA- or SemiAnalysis-run numbers (open-source signal): The preview results do not reproduce outside the vendor-assisted setup.

Still open

  • Will non-preview Rubin results on mature vLLM, SGLang and TRT-LLM stacks change the roughly 2x mainstream ratio?
  • Has SemiAnalysis completed its review of NVIDIA's self-measured 30x result?
  • How well does a benchmark built only from Claude Code sessions, with no evaluation of answer quality, predict other agentic workloads?
  • Why do pages on the official InferenceX host carry an 'unaffiliated copy' disclaimer?

Evidence / sources

  1. 1inferencex.semianalysis.com/hardware/vera-rubin-nvl72-vs-gb300-nvl72-agentx-results
  2. 2developer.nvidia.com/blog/nvidia-vera-rubin-and-blackwell-set-a-new-standard-for-agentic-ai-performance-per-watt
  3. 3newsletter.semianalysis.com/p/vera-rubin-nvl72-agentic-inference
  4. 4inferencex.semianalysis.com/blog/vera-rubin-nvl72-agentic-inference
  5. 5inferencex.semianalysis.com/agentx/methodology
  6. 6nvidia.com/en-us/data-center/vera-rubin-nvl72
  7. 7github.com/SemiAnalysisAI/InferenceX
  8. 8inferencex.semianalysis.com/about
All Analysis

Next reads

Read next

Open the graph
  1. 01topicOpen SourceExplore the open_source topic hub.
  2. 02archiveAnalysisOpen the latest authorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract